Together is a GPU reseller

The important point is that Together sits one layer above the hardest part of the stack, which is owning scarce, expensive GPU clusters, and that shapes both its appeal and its limits. Lambda and CoreWeave buy or control the hardware, reserve power, and sign long contracts. Together buys capacity from them and repackages it into a simpler product for startups that want model training and inference without committing to fixed GPU blocks or building infra teams.

• Lambda and CoreWeave sell reserved access to networked GPU clusters, often for 12 to 18 month terms, because training large models needs identical GPUs, fast interconnects like InfiniBand, and guaranteed uptime. That is infrastructure ownership, not brokerage, and it is why they can serve the heaviest training workloads directly.
  2 sacra 3 sacra 5 sacra
• Together changes the buying model. Instead of asking a startup to reserve a whole cluster, it charges in a usage based way and bundles open model access, training tools, and inference APIs. That makes it attractive for smaller customers with spiky demand, but it also means gross margins are structurally lower than the underlying infrastructure providers.
  1 sacra 3 sacra 4 sacra
• The market is already splitting by customer type. CoreWeave has moved upmarket into very large enterprise contracts. Lambda is the more flexible training cloud for growth stage teams. Together wins when the customer mostly wants convenience, model compatibility, and not having idle GPUs on the balance sheet.
  1 sacra 2 sacra 4 sacra 5 sacra

Going forward, this stack separation should get sharper. Infrastructure owners will keep consolidating around capital, chip access, and power, while companies like Together will live or die on software, routing, and developer experience. As GPU supply loosens, the resellers with the best abstraction layer should gain leverage, while the weakest will be squeezed by falling infrastructure prices and copyable APIs.